Reasoning

In short: Ask for step-by-step reasoning BEFORE the answer.

What is: Have the model write its reasoning in steps before concluding, instead of jumping straight to the answer. It's the single most important reasoning technique.

When to use: Math, logic, decisions with trade-offs, hard classification.

When to avoid: On reasoning models (OpenAI o-series, GPT-5 reasoning): they already reason internally and asking for "step by step" can even hurt. On Claude with thinking off, prefer "analyze/assess", the word "think" is sensitive.

💡 Tip: Separate the reasoning (<thinking>) from the final answer (<answer>) so you use the thinking without polluting the output.

In short: Just add "think step by step", no examples.

What is: The simplest version of CoT: append a phrase like "Let's think step by step" without providing reasoning examples.

When to use: When you want the CoT gain without building examples.

When to avoid: On reasoning models (they reason on their own). On Claude without thinking, swap "think" for "consider/analyze".

💡 Tip: It's practically free. If the task is even slightly analytical, it's always worth a try.

In short: Step back to a general principle before solving the case.

What is: Before attacking the specific problem, ask the model to state the general concept or rule. Then it applies that rule to the case.

When to use: Problems where applying the right rule matters more than computing, physics, law, diagnosis.

When to avoid: On direct, simple tasks it adds steps with no gain.

💡 Tip: Combine with CoT: state the principle, then reason step by step on the application.

In short: Plan first, execute after.

What is: An improvement on "think step by step": ask the model to FIRST draft a plan and THEN execute. Reduces the error of skipping steps.

When to use: Multi-step tasks where simple "step by step" sometimes forgets parts.

When to avoid: Single-step tasks, the planning becomes bureaucracy.

💡 Tip: Ask for the plan as a list before execution, easy to review and course-correct.

In short: Solve several times and go with the majority answer.

What is: Run the same CoT prompt several times (with randomness) and pick the answer that appears most often. Like polling several experts and following the consensus.

When to use: Critical problems where a single chain can slip and you can afford N runs.

When to avoid: Simple or cost-sensitive tasks, it costs N× tokens.

💡 Tip: Even in a single call, ask for "3 independent approaches + majority vote" for a similar effect at lower cost.

In short: Ask the model to check its own answer before delivering.

What is: Add a step where the model verifies the answer against a criterion before finishing. Cheap and effective.

When to use: Almost always when the answer is verifiable, code, math, extraction.

When to avoid: When there's no objective verification criterion, it's just more text.

💡 Tip: Give explicit verification criteria ("test with X, Y, Z"), vague verification catches little.

In short: Generate, critique your own answer, and redo it better.

What is: A loop: the model generates, CRITIQUES its own answer pointing out concrete flaws, then produces a revised version. Stronger than self-check (which is just verification).

When to use: Code, writing, sequential reasoning, tasks where the 1st try usually has fixable errors.

When to avoid: When the first answer is already good and simple, the loop only adds cost.

💡 Tip: Give concrete criteria for the critique; a generic "improve it" yields cosmetic changes.